Re: [PATCH 1/6] block: add support for carrying a stream ID in a bio

From: Jens Axboe <axboe@kernel.dk>
To: "Ming Lin-SSI" <ming.l@ssi.samsung.com>,
	"Matias Bjørling" <m@bjorling.me>, "Jens Axboe" <axboe@fb.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>
Subject: Re: [PATCH 1/6] block: add support for carrying a stream ID in a bio
Date: Tue, 24 Mar 2015 19:42:21 -0600	[thread overview]
Message-ID: <5512127D.2040508@kernel.dk> (raw)
In-Reply-To: <3A47B4705F6BE24CBB43C61AA73286215067DA@SSIEXCH-MB3.ssi.samsung.com>

On 03/24/2015 04:07 PM, Ming Lin-SSI wrote:
>> -----Original Message-----
>> From: Jens Axboe [mailto:axboe@kernel.dk]
>> Sent: Tuesday, March 24, 2015 10:27 AM
>> To: Matias Bjørling; Jens Axboe; linux-kernel@vger.kernel.org; linux-
>> fsdevel@vger.kernel.org
>> Cc: Ming Lin-SSI
>> Subject: Re: [PATCH 1/6] block: add support for carrying a stream ID in a bio
>>
>> On 03/24/2015 11:11 AM, Matias Bjørling wrote:
>>> On 03/24/2015 04:26 PM, Jens Axboe wrote:
>>>> The top bits of bio->bi_flags are reserved for keeping the allocation
>>>> pool, set aside the next four bits for carrying a stream ID. That
>>>> leaves us with support for 15 streams,
>>>> 0 is reserved as a "stream not set" value.
>>>
>>> 15 streams seem very limited. Can this be extended? e.g. 16 bits.
>>>
>>> 15 streams is enough for 1-4 applications. More, and applications
>>> starts to fight over the same stream id's, leading them to place
>>> different age data in same flash blocks and push us back to square one.
>>>
>>> I understand that Samsung multi-stream SSD supports a limited amount
>>> of streams, more advance implementations should provide higher limits.
>>
>> Pushing it higher is not a big deal as far as the implementation goes, though
>> 16 bits might be stealing a bit too much space for this. On 32-bit archs, we
>> have 18 bits currently free that we can abuse. The Samsung device supports
>> 16 streams. That's honestly a lot more than I would expect most devices to
>> support in hardware, 16 is a lot of open erase blocks and write append points.
>> Obviously the open channel effort would make that more feasible, though.
>
> Can we use 8 bits at least? I'll test performance with 16 streams.

We could, but I still question whether that's really useful. I'd rather 
start smaller and go bigger if there's a real use case for it. It wont 
change the user space ABI if we later make it larger.

-- 
Jens Axboe